core Estimated learning time: 5 h

1.17 Maximum likelihood and MAP

You can derive a loss function instead of memorizing it.

Before:00. Orientation & SetupUnlocks:03. Data Handling & Analysis04. Classical AI — Agents, Search & Knowledge Representation

Maximum likelihood is the recipe that generates most loss functions: write the probability of the data, take the negative log, minimise. Once cross-entropy falls out of a Bernoulli likelihood in front of you, losses stop being arbitrary facts to memorise. MAP adds a prior and lands exactly where regularisation will. It closes the estimation arc because deriving two losses once replaces memorising twenty. The confusion to fix early is likelihood versus probability — same formula, different variable held fixed.

Work through these

  • Likelihood vs. probability

    Probability asks how likely data is given a model; likelihood asks how well a model explains data already seen. Swapping the two is a common and consequential confusion.

    Introduction to Probability & Statistics (18.05) · Course
  • MLE for Bernoulli, normal, linear models

    Deriving the best-fitting parameters for the standard simple models, which is where loss functions come from rather than being handed down. Doing it once changes how you read every later loss.

  • Priors, posteriors and MAP estimation

    Adding a prior belief and finding the most probable parameters afterwards, which is the smallest step from one estimation philosophy to the other. It also explains certain penalty terms.

  • Why cross-entropy is just negative log-likelihood

    The classification loss used everywhere in deep learning is exactly a negative log-likelihood. Seeing that removes the sense that loss functions are arbitrary choices.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.